Stanford's Matryoshka Attribution Tops Mechanistic Interpretability Benchmark at 2.9x Runner-up

On 09-24, a Stanford team (including Aryaman, Chris Potts, and Dan Jurafsky) released Matryoshka Attribution (MAttr), a gradient-descent-based attribution method for identifying which parts of a network are responsible for a behavior, scoring 2.9x the runner-up on mechanistic interpretability benchmarks.

Confirmed

• Positioning: The authors frame classic interpretability problems such as attribution as training objectives optimized directly via gradients, explicitly responding to the "Impossibility Theorem" paper by Bilodeau, Jaques, Koh, and Kim.
• Technique: MAttr combines causal interventions, gradient attribution, and mask learning, parameterizing interventions with a simple differentiable sigmoid top-k operator—no sparsity losses or straight-through tricks. Randomizing the top-k budget during training induces an attribution ranking, learns nested orderings of model components across sparsity levels, and optimizes intervention performance directly.
• Results: It scores 2.9x the runner-up on mechanistic interpretability benchmarks; the authors report SOTA after roughly 500 training steps, using judge scores as reward.
• Application: Beyond representation-level attribution, MAttr traces an LLM's qualitative behavior to specific parameter subsets updated during training—using StrongREJECT scores as RL reward for attribution and applying the mask to weights between two checkpoints (not representations between two prompts), localizing Llama 3.1's refusal behavior to 1% of the parameter delta.

Why it matters

• The approach sidesteps theoretical objections to axiomatic attribution raised by the Impossibility Theorem, offering an optimizable engineering path for interpretability research.
• By localizing high-level behaviors such as refusal to a tiny fraction of parameters, it provides a practical tool for understanding and editing large-model behavior (e.g., removing refusals).

2026-09-24 ~ 2026-09-24 · 7 related posts

• [source] Matryoshka Attribution tops interpretability benchmark by 2.9x, traces LLM refusals to 1% of weights — aryaman2020 · 2026-09-24
• [source] Stanford's Matryoshka Attribution tops mech interp benchmark at 2.9x the runner-up — aryaman2020 · 2026-09-24
• MAttr technical details: differentiable sigmoid top-k merges causal interventions and mask learning — aryaman2020 · 2026-09-24
• MAttr turns attribution into a training objective, hitting MIB SOTA in just 500 steps — aryaman2020 · 2026-09-24
• Attribution as a training objective: MAttr answers the Impossibility Theorems paper — aryaman2020 · 2026-09-24
• [source] MAttr traces Llama 3.1 refusals to 1% of parameter delta, resetting them disables refusal — aryaman2020 · 2026-09-24

1 near-duplicate retellings: ChrisGPotts